Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

ACCELERATING MULTISCALE MATERIALS MODELING WITH MACHINE LEARNING. Multiscale materials modeling provides

in electron density of the new surrogate model is under 1%, but computation is three orders of magnitude faster. Promising methodologies such as optimal experimental design techniques and novel Graph Neural Networks were explored in training smaller data sets and will be researched further in the future to continue accelerating first-principle data generation and increase the fidelity and robustness of predictive atomistic materials simulations. An ML model designed for aluminum has already been successfully leveraged in Sandia’s Electronics Parts Program milestone. (PI: Sivasankaran Rajamanickam)

fundamental insight into microscopic mechanisms that determine materials properties in nuclear stockpile applications that leverage radiation harden semiconductors, advanced manufacturing, shock compression, and energetic materials. This LDRD team including three postdoctoral researchers developed a new ML surrogate model for density functional theory using deep neural networks to accurately predict total energies of 100,000 atom systems when trained on only 256 atoms. When compared with direct numerical simulation of 2048 aluminum atoms, the error

OPTIMIZING MACHINE LEARNING DECISIONS WITH PREDICTION UNCERTAINTY. While ML classifiers are widespread, their

demonstrations on cyber security and image analysis cases. The developed and trained ML classifier ultimately provided a framework for: (1) quantifying and propagating uncertainty in ML classifiers; (2) formally linking ML outputs with the decision-making process; and (3) making optimal decisions for classification under uncertainty with single or multiple objectives. Methods developed through this project are currently being incorporated into applications that impact national security domains by directly addressing questions of automated decision. (PI: Michael Darling)

output is often not part of a follow-on decision making process because of lack of uncertainty quantification. Through this LDRD, the team developed decision analysis methods that combined uncertainty estimates for ML predictions with a domain-specific model of error costs. In the project, they explicitly weighed whether ML models under evaluation were qualified to make any prediction by producing general algorithms that minimized prediction error costs by validating these algorithms through

67

LABORATORY DIRECTED RESEARCH & DEVELOPMENT